Learned-guidance search

Let a small learned board evaluator decide where a deeper search should spend its time, while keeping the exact search as a safety net.

The intuition

A complete look-ahead spends the same effort on every column, including the ones a competent player would dismiss instantly. If something could tell the search which branches are worth expanding, the same budget would reach further ahead. That "something" here is a small learned evaluator of the public position — an NNUE-style network trained to reproduce a stronger search's judgements.

The design is deliberately timid, and that is its most interesting feature. The learned model is never allowed to be the last word:

• Exact first. Complete the full-width sampled-chance depth-3 search to a finished answer. Only then spend whatever budget remains on learned-guided deeper iterations, and only accept a deeper answer if that whole iteration finished.
• Root actions are untouchable. The model may order and prune interior branches; every legal column at the current position is always searched.

Three programs sit in this directory. Two of them test guidance; the third tests the opposite lever — spending the extra budget on covering more chance outcomes at the root instead of on more depth — using the same fixed model and feature code so the comparison is clean.

How it works, step by step

1. Read the public position and run the exact depth-3 search to completion. Record its answer.
2. With the remaining fixed work budget, run deeper iterations in which the learned evaluator orders and reduces interior choice nodes.
3. If a deeper iteration completes within budget, and only then, allow its answer to replace the depth-3 answer.
4. Play that column.

The root-quadrature variant replaces step 2 with a wider treatment of chance: three first-reveal strata crossed with all seven possible next discs, followed by three joint interior strata.

What happened, in plain English

Every version was worse than the plain depth-3 search it was trying to improve, and the ones that spent the most extra work were the worst.

The safeguarded deeper search spent nearly its whole extra budget, actually completed a deeper iteration on about a fifth of moves, and changed the depth-3 answer on fewer than one move in twenty — and still came out behind on score and on how long the games lasted. That combination is the finding: a rare learned override is not harmless, because the rare occasions on which a model disagrees with an exact search are exactly the occasions where its errors are largest. The wider chance coverage at the root failed the same way for a different reason — it spent about 239,000 work units per move against about 58,000 for plain depth 3 and still lost to it by a wide margin, showing that better sampling of chance does not repair an evaluator that is biased at the horizon.

The third program, which re-searched only the actions the model liked, is recorded as rejected but has no entry in the written ledger at all. Its number exists only in a research conversation, so there is nothing here to quote.

Two limitations must travel with all of this. These screens are four games each, which is small even for a screen. And they were decided under the historical 7,000-point Sequence scoring, which the repository treats as archival: the level bonus is now 17,000 points and carries the overwhelming majority of a Hardcore score, so a result that traded score against survival could plausibly change sign under the corrected rules. No corrected re-run of any of these three is on record.

The technical record

Learned deeper-search override — rejected, ledger-recorded. On four historical 7,000-point Sequence-scored games, the safeguarded policy averaged 82,169.25 points and 58.75 moves versus 87,743.75 and 62 for depth 3. It spent 235,654 work units per move, completed a deeper iteration on 22.1% of moves, and changed the depth-3 action on 4.68% of moves. Because even the screen mean was worse, the reserved confirmation seeds were not read.

Root reveal quadrature — rejected, ledger-recorded. Its predeclared absolute gate was the full-width sampled-chance depth-3 policy. On four historical 7,000-point Sequence-scored games, depth 3 averaged 205,001.25 points and 132.5 moves while the larger quadrature averaged 179,243 and 120. It used about 239,000 work units per move against about 58,000, failed the screen, and confirmation was not run.

Selective learned search — rejected, task-record only. There is no ledger protocol, no artifact hash and no retained per-game data; audit 04 lists this source among the thirty rejections that "cannot be re-derived". From the source it re-searches only model-ranked promising actions while retaining the frozen exact fallback.

A caveat on "fresh" seeds. The ledger describes the seed ranges for the deeper-search and quadrature screens as fresh. Audit 03 — an exploratory, read-only audit, not a re-measurement — finds those ranges fall inside the first 100,000-game n-tuple training block, at offsets 45,056 and 49,152, and places the selective-search screen and confirmation ranges at offset 36,864 of the same block. That does not change these rejections, which are against a comparator on the same games, but it means the games were not new to the repository.

Scoring. Unlike their observable-MCTS neighbours, these three sources carry no compile-time 7,000-point lock, so they build against today's corrected 17,000-point engine. Re-running them therefore would not reproduce the numbers above.

Sources: "Learned deeper-search override (rejected)" and "Root reveal quadrature (rejected)" in the experiment history; the tree-search table in the experiment index.

What this taught us, and what is still open

• A rare override is a risky override. Changing under 5% of decisions still lost the game, because a model disagrees with an exact search precisely where it is least reliable. The same shape appears again in the terminal policy iteration family.
• More chance coverage is not a substitute for a better leaf. The quadrature arm isolated that lever and it did not help.
• The current guidance is explicit about ordering. The research roadmap states: "Do not begin with learned action pruning. Every legal root action remains eligible until the model has independently demonstrated safe sibling ranking." These results are why.
• What was ruled out is these configurations under old scoring on four-game screens; what was not ruled out is learned guidance in general, which remains untested under corrected Hardcore scoring on a cohort large enough to decide anything.